Program

 

09:00

09:15

 

Workshop Opening

09:15

10:00

Invited Speaker – Prof. Jennifer Coates

10:00

12:30

Oral Session 1

10:00

10:15

Food/Non-food Image Classification and Food Categorization using Pre-Trained GoogLeNet Model

Ashutosh Singla, EPFL; Lin Yuan, EPFL; Touradj Ebrahimi, EPFL

10:15

10:30

Foodness Proposal for Webly-supervised food detection

Wataru Shimoda, The University of Electro-Communications, Tokyo; Keiji Yanai, Univ. Electro-Comm., Tokyo

10:30

10:45

Food Image Segmentation for Dietary Assessment

Joachim Dehais, University of Bern; Marios Anthimopoulos, University ofBern, Switzerland; Stavroula Mougiakakou, University of Bern, Switzerland

10:45

11:15

Coffee Break

11:15

12:00

Invited Speaker – Prof. Efstratios Gavves

12:00

12:30

Oral Session 2

12:00

12:15

Snap, eat, repEat: a Food Recognition Engine for Dietary Logging

Michele Merler, IBM Research; Hui Wu, IBM Research; Rosario Uceda-Sosa, IBM Research; Quoc-Bao Nguyen, IBM Research; John Smith , IBM

12:15

12:30

Food Image Recognition with Very Deep Convolutional Networks

HAMID HASSANNEJAD, University of Parma; Guido Matrella, University ofParma; Paolo Ciampolini, University of Parma; Ilaria De Munari, University of Parma; Monica Mordonini, University of Parma; Stefano Cagnoni, University of Parma

12:30

14:00

Lunch Break

14:00

14:45

Invited Speaker – Dr. Thomas Mensink

14:45

15:30

Oral Session 3

14:45

15:00

A Mobile Food Record For Integrated Dietary Assessment

Ziad Ahmad, Purdue University

15:00

15:15

An Automatic Calorie Estimation System of Food Images on a Smartphone

Koichi Okamoto, The University of Electro-Communications, Tokyo; Keiji Yanai, Univ. Electro-Comm., Tokyo

15:15

15:30

Deep-based Ingredient Recognition for Cooking Recipe Retrieval (INVITED)
Jingjing Chen, City university of Hong Kong; Chong-Wah Ngo, City Universityof Hong Kong

15:30

16:00

Coffee Break

16:00

17:30

Poster Session


Food/Non-food Image Classification and Food Categorization using Pre-Trained GoogLeNet Model

Ashutosh Singla, EPFL; Lin Yuan, EPFL; Touradj Ebrahimi, EPFL

Foodness Proposal for Webly-supervised food detection

Wataru Shimoda, The University of Electro-Communications, Tokyo; Keiji Yanai, Univ. Electro-Comm., Tokyo

Food Image Segmentation for Dietary Assessment

Joachim Dehais, University of Bern; Marios Anthimopoulos, University of Bern, Switzerland; Stavroula Mougiakakou, University of Bern, Switzerlan

Snap, eat, repEat: a Food Recognition Engine for Dietary Logging

Michele Merler, IBM Research; Hui Wu, IBM Research; Rosario Uceda-Sosa, IBM Research; Quoc-Bao Nguyen, IBM Research; John Smith , IBM

Food Image Recognition with Very Deep Convolutional Networks

HAMID HASSANNEJAD, University of Parma; Guido Matrella, University of Parma; Paolo Ciampolini, University of Parma; Ilaria De Munari, University of Parma; Monica Mordonini, University of Parma; Stefano Cagnoni, University of Parma

A Mobile Food Record For Integrated Dietary Assessment

Ziad Ahmad, Purdue University

An Automatic Calorie Estimation System of Food Images on a Smartphone

Koichi Okamoto, The University of Electro-Communications, Tokyo; Keiji Yanai, Univ. Electro-Comm., Tokyo

Food Search Based on User Feedback to Assist Image-based Food Recording Systems

Sosuke Amano, The University of Tokyo; Shota Horiguchi, The University of Tokyo; Kiyoharu Aizawa, The University of Tokyo; Kazuki Maeda, foo.log Inc.; Masanori Kubota, foo.log Inc.; Makoto Ogawa, foo.log Inc.

Food vs Non-Food Classification

Francesco Ragusa, University of Catania; Valeria Tomaselli, STMicroelectronics; Antonino Furnari, University of Catania; Sebastiano Battiato, Università di Catania; Prof. Giovanni Maria Farinella, “University of Catania, Italy – CV”

Performance Evaluation Methods of Computer Vision Systems for Meal Assessment

Marios Anthimopoulos, University of Bern, Switzerland; Joachim Dehais, University of Bern; Stavroula Mougiakakou, University of Bern, Switzerland

Invited

Learning to Make Better Mistakes: Semantics-aware Visual Food Recognition
Hui Wu, IBM Watson Research; Michele Michele Merler, IBM TJ Watson Research Center; Rosario Uceda-Sosa, IBM Research; John Smith, IBM Research

Invited

Deep-based Ingredient Recognition for Cooking Recipe Retrieval
Jingjing Chen, City university of Hong Kong; Chong-Wah Ngo, City University of Hong Kong

16:00

17:30

Demo Session

DeepFoodCam: A DCNN-based Real-time Mobile Food Recognition System

Ryosuke Tanno, The University of Electro-Communications, Tokyo; Koichi Okamoto, The University of Electro-Communications, Tokyo; Keiji Yanai, Univ. Electro-Comm., Tokyo

Abstract: Due to the recent progress of the studies on deep learning, deep  convolutional neural network (DCNN) based methods have outperformed conventional methods with a large margin. Therefore, DCNN-based recognition should be introduced into mobile object recognition. However, since DCNN computation is usually performed on GPU-equipped PCs, it is not easy for mobile devices where memory and computational power is limited. In this demo, we show the possibility of DCNN-based object recognition on mobile devices, especially on iOS and Android devices including smartphones and tablets in terms of processing speed and required memory. As an example of DCNN-based mobile applications, we show a food image recognition app, “DeepFoodCam”. In this demo, we show DCNN based “DeepFoodCam” outperformed FV-based “FoodCam” greatly in terms of recognition accuracy

 We use multi-scale network-in-networks (NIN) in which users can adjust the trade-off between recognition time and accuracy. We implemented multi-threaded mobile applications on both iOS and Android employing either NEON SIMD instructions or the BLAS library for fast computation of convolutional layers, and compared them in terms of recognition time on mobile devices. As results, it has been revealed that BLAS is better for iOS, while NEON is better for Android, and that reducing the size of an input image by resizing is very effective for speedup of DCNN-based recognition. In case of using iPad Pro with BLAS, we achived 66.0ms as the processing time for one time recognition.

GoCARB: a Smartphone Application for Automatic Assessment of Carbohydrate Intake

Joachim Dehais, University of Bern; Marios Anthimopoulos, University of Bern, Switzerland; Stavroula Mougiakakou, University of Bern, Switzerland

Abstact: Dietary and lifestyle management rely on objective and accurate diet assessment. To assess dietary intake itself requires training and skills however. In that regard, trained individuals often misjudge what they eat, even when they are under strict constraints. These issues emphasize the need for objective, accurate dietary assessment tools that can be delivered to and directly used by the public to monitor their intake.

We demonstrate such a tool adapted to the specific assessment needs of individuals with type 1 diabetes mellitus. The tool itself is a combination of a smartphone  application, and a processing server, designed to calculate the amount of carbohydrate within a dish. This overall system requests the user to take a pair of images of the food from different points of view using the smartphone application, after which the data is transferred to the server, analyzed, and the results sent back to the user for confirmation. The application itself guides the user through the capture and the review of results, while the server does all the data processing.

The first processing step operates by initially detecting the dish, using  a fast RANSAC detector, and then separate it from its content, using efficient region growing and merging operations. This step shows accuracy levels of nearly 88%. The result of this step is sent to the user who can choose to either accept the result or correct it using a semi-automatic segmentation interface. The second step is to recognize the content, which is done using both color and texture features and SVM classification. The  resulting classification confidences are then used to organize a list of top choices and let the user confirm the top choice or select the right one. Finally, the images, segmentation, and classes are used to build 3D models of the food items, calculate their volume, and from that their nutrient content. The entire procedure lasts for about 7 seconds, without counting the user interaction, and the average error in carbohydrate estimation was found in the order of 10% in a technical evaluation. GoCARB is one of the first complete meal assessment systems.

Industry

MealoMi: A mobile application for automated food assessment trained on RBG-D food data using Residual Fully Convolutional Neural networks

Sebastian Schlecht, MealoMi.com; Florian Ettlinger, MealoMi.com; Patrick Ferdinand Christ, Technical University of Munich

Abstract: Digital food assessment is a field gaining increasing interest in recent years. With the advances of mobile technologies, it became easier and more convenient to log and track dietary behavior of individuals via a multitude of applications. Most of these applications however miss the part of automated food assessment, i.e. capturing the kind and amount of food consumed during the meal, leading to lower retention rates for these systems. At MealoMi App, we automatically classify meals from food images and investigate depth prediction for meals in order to be able to estimate volumes inside the scene later. The depth is inferred using very deep residual fully convolutional neural networks which are supposed to be trained on RGB-D image pairs. This depth-map is used in our algorithms to predict the actual volume of the objects. To support specific depth prediction for meals we modified a Microsoft Kinect in order to collect RGB-D food data to train our deep learning algorithms.

17:30

18:00

Panel discussion